acceptodds
Under review as a conference paper at ICLR 2027

STOCHASTIC ODE-GUARD: ANTI-CONCENTRATION BOUNDS FOR NEURAL SDE CLASSIFIERS WITH APPLICATION TO ADVERSARIALLY ROBUST INTRUSION DETECTION

Abstract

Neural intrusion detectors classify network flows from a few dozen traffic statistics. They lose much of their accuracy when an attacker nudges those statistics slightly, and the defences that come with formal guarantees either treat the classifier as a fixed function or add noise to its input. We study Stochastic ODE-Guard (SODE-Guard), a neural stochastic differential equation (NSDE) classifier: it maps a flow to an internal representation, lets that representation evolve under a learned random dynamics, and classifies by averaging over the random outcomes. We prove two robustness results for this model class. (i) Theorem A: the averaged classifier, exactly as it is computed with the Euler–Maruyama scheme, changes by at most L times the size of any input change, with L computed from the trained weights. A score gap M(x) between the top two classes therefore guarantees the same decision for every input change smaller than M(x)/(√2 L); a Monte-Carlo confidence correction makes the guarantee usable with finitely many samples. (ii) Proposition 2: for a single random outcome and any fixed input change, the probability that the decision flips is at most the probability of starting within β of the decision boundary plus 2L²ε²/β²; our anti-concentration regulariser is designed to shrink the first term. We also state exactly when a Bismut–Elworthy–Li estimator recovers input gradients of such a model. On three public flow benchmarks SODE-Guard attains 96.4% clean macro-F1 and 93.1% under PGD-40 at ε = 0.03, against 87.2% for the strongest neural baseline, at 1.7 ms median latency per batch of 128 on one A100.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
adversarial robustness, certified robustness, neural stochastic differential equations, anti-concentration inequalities, uncertainty quantification, Malliavin calculus, network intrusion detection

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_kdckOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SODE-Guard for intrusion detection, evolving latent representations through a neural SDE and averaging predictions over random trajectories. It derives robustness bounds for the averaged predictor and individual paths, and evaluates predictive performance, ablations, and efficiency on three datasets.

Critical strengths

1.The robustness analysis distinguishes averaged predictions from individual trajectories and addresses the discretised SDE model. 2.Ablations examine diffusion, spectral normalisation, multi-path inference, and regularisation, helping assess individual components.

Critical weaknesses

1.Insufficient design motivation. The introduction outlines the method but does not adequately explain how latent stochastic dynamics and anti-concentration regularisation address specific limitations of existing approaches. 2.Figure quality and proofreading. Figure 1 is crowded, with some text extending beyond boxes. Figures 2(a) and 2(b) have inconsistent plotting dimensions and alignment. The figures should be redrawn or carefully revised for readability. 3.Limited baseline coverage. Several baselines date from 2022. Although IDS-GraphMamba (2025) is included, comparisons should also cover recent methods directly targeting adversarial or certified robustness under the same threat model. 4.Disconnect between certification theory and experiments. Section 4 explicitly states that the evaluated encoder does not satisfy Assumption 1, and no certified radii are reported. The central certification contribution therefore remains unvalidated on the experimental models. 5.Incomplete adaptive attack evaluation. Results mainly rely on PGD-40 against the eight-path average. Larger-budget EOT and attack ensembles are not reported, leaving insufficient evidence to rule out gradient masking and substantiate the empirical robustness claims.

Questions for authors

1.Are random trajectories resampled during PGD? Are results available for larger EOT budgets or other adaptive attacks?

1 comment
AuthorsOpenReview ↗

Response to reviewer kdck: Update after the five-seed re-run (see also the general update)

Thank you again for the constructive review. Here is how the update answers each point.

Q1 (paths during PGD). Every attack is scored under both settings:

  • fixed Brownian paths (seeds 0-7, a deterministic target, attacked exactly by PGD with restarts);
  • fresh paths on every call (EOT-PGD, APGD with EOT, and a 64-path expected-margin attack).

The two agree closely. Worst-case macro-F1 for SODE-Guard with EOT-matched adversarial training is 0.420 with fixed paths and 0.412 with fresh paths.

W5 (adaptive attacks). Adversarial training through the deployed 8-path predictor, with fresh paths in every step (module 2), raises worst-case macro-F1 from 0.245 to 0.420 ± 0.010 over five seeds (Holm-adjusted p = 0.016). That is on par with an adversarially trained MLP (0.424) and neural ODE (0.432); the differences are not significant. The masking checks pass. A full-strength re-check of seed 0 (PGD-200 with 5 restarts, EOT-32, a 256-path expected-margin attack, Square with 5,000 queries) is still running now.

W4 (certification). SODE-Guard is now certified with randomised smoothing over the input and its own Brownian paths. Certified accuracy at l2 radius 0.273 is 0.556 ± 0.033, against 0.096 with Theorem A before. That is on par with a smoothed MLP (0.555). Reweighted to the test class mix, it is 0.843.

W3 (recent and strong baselines). We added:

  • LightGBM, XGBoost and a random forest;
  • adversarially trained MLP and neural ODE, with the same recipe as SODE-Guard;
  • a smoothed MLP;
  • a neural ODE with the same imbalance-aware training.

The trees reach the highest clean macro-F1 (0.61-0.62) but fall to 0.02-0.04 under a black-box Square attack, while adversarially trained neural models keep about 0.42. The revision discusses BARS (NDSS 2023) and MARS (2025) as certified traffic classifiers with anisotropic smoothing, a natural extension of module 3.

W1, W2 (motivation, figures). The robustness gap of the strongest clean models gives the revision a concrete motivation:

  • trees are best on clean flows and weakest under attack;
  • a robust neural detector closes that gap.

Figures 1-2 will be redrawn around the three modules and the clean / worst-case / certified trade-off.

Running now, for the revision:

  • per-module ablations over five seeds;
  • a per-class chronological split;
  • the full-strength attack check;
  • also planned: feature-constrained attacks and a second container dataset.

We hope the update addresses your concerns, and would be grateful if you could consider it in your final assessment. Thank you.

Reviewer_FiAvOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

SODE-Guard is a neural SDE classifier for network intrusion detection. It encodes flow statistics into a latent state, evolves the state with a discretized SDE, and classifies with a linear head averaged over several random paths. The theory gives three results: a Lipschitz-based certified radius for the path-averaged predictor; a bound on single-path decision flips ; a regularized Bismut-Elworthy-Li gradient estimator. On three flow benchmarks, it reports better macro-F1 under PGD-40 than six neural baselines, as well as the best clean macro-F1.

Critical strengths
  • The paper is precise about its scope. The bounds are stated for the discretized scheme that is actually run, and the paper says clearly what is and is not certified.

  • The empirical results under the evaluated attack come with good calibration and low latency.

Critical weaknesses
  • The main contribution is not validated. The evaluated models do not satisfy Assumption 1, and the authors report no certified radii for them. The title promises certified bounds, yet the paper reports no certified accuracy. It also has no comparison with certified baselines such as randomized smoothing or Lipschitz-bounded networks.

  • The theory is standard and does not support the design. Each result follows from well-known tools: mean-square stability, the Lipschitz-margin certificate of Tsuzuku et al., Hoeffding or Bernstein bounds, Markov’s inequality and the Elworthy-Li formula. Theorem A also shows that the diffusion enlarges the worst-case constant. The analysis therefore gives no reason to prefer an SDE over an ODE for certification, as Remark 2 admits. The closest prior work, Liu et al. (CVPR 2020) on how noise in neural SDEs improves robustness, is not cited.

  • The robustness may come from gradient masking. The only attack is PGD-40 through the path average, without large-sample EOT, AutoAttack or a query-based attack. Removing the diffusion hurts robust accuracy far more than clean accuracy, a typical pattern of stochastic gradient obfuscation. The authors themselves call these numbers upper bounds until adaptive attacks are run.

  • The intrusion-detection protocol is weak. The datasets are anonymized compilations, and results use random splits although temporal and host-disjoint splits exist in the repository. All methods reach about 0.95 clean macro-F1, which suggests saturated or leaky splits (see Arp et al., USENIX Security 2022). The attacks also perturb normalized features without the constraints of real traffic.

Questions for authors

Please answer the questions raised in my review.

1 comment
AuthorsOpenReview ↗

Update after the five-seed re-run (see also the general update)

Thank you again. The update responds directly to your four points.

Contribution not validated

  • What we now certify: a trained SODE-Guard, under a randomised-smoothing certificate taken jointly over input noise and the model's own Brownian paths. The input noise is a pure-diffusion prefix of the SDE.
  • Result: certified accuracy at l2 radius 0.273 (which contains the l-inf 0.03 ball) is 0.556 ± 0.033 over five seeds, against 0.096 for Theorem A in our first reply.
  • Against baselines: this is on par with a smoothed MLP (0.555, difference not significant).
  • Reweighted to the test class mix, the certified accuracy is 0.843.

Why an SDE We agree that the learned diffusion does not, by itself, add accuracy or robustness:

  • SODE-Guard vs a neural ODE: +0.005 clean macro-F1 and -0.005 worst-case macro-F1, are both not significant.

But, what the stochastic model offers is a natural certificate. The deployed predictor already averages over random paths, so smoothing over input noise and those same paths needs no separate noisy copy of the model. The revision presents the contribution accordingly: a training-and-certification recipe for stochastic continuous-depth detectors that brings SODE-Guard level with the strongest robust baselines. We also cite and discuss Liu et al. (CVPR 2020), as you suggested.

Gradient masking With EOT-matched adversarial training, worst-case macro-F1 over eight adaptive attacks is:

  • 0.420 with fixed paths;
  • 0.412 with fresh paths;
  • 0.245 without adversarial training.

The checks show no masking:

  • F1 falls to 0.004 at eps = 1.
  • More steps do not help the model.
  • Square and transfer are weaker than the white-box attacks.

Protocol (Arp et al.)

  • Duplicates and conflicting labels: we removed this before splitting (325 exact duplicates and 1,250 flows whose features appear with a different label).
  • Seeds and statistics: five seeds with new splits, paired bootstrap with Holm correction, and per-class F1.
  • Baselines: we added gradient-boosted trees and a random forest. They reach the highest clean macro-F1 (0.61-0.62) but fall to 0.02-0.04 under a black-box Square attack, while adversarially trained neural models, SODE-Guard included, keep 0.42.
  • Rare classes: two of them are poorly separated by every model, trees included. The revision analyses this as a property of the data.
  • Temporal evaluation: a per-class chronological split replaces the global temporal split, whose test window held one class. It is running now.
  • Realistic perturbations: feature-constrained attacks are planned for the revision.

We hope this resolves your main concerns, and thank you for the references and suggestions that led to these results. Thank you.

Reviewer_vjJKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies a neural SDE classifier for intrusion detection. It evolves an encoded flow through noisy latent dynamics and averages class scores. The main theory gives a finite-step robustness certificate for an expected projected-score classifier, with a Monte Carlo correction. The paper also gives a single-path bound and proposes a margin regularizer and BEL input gradients. A related trained model is compared with six detectors on 3 datasets under PGD-40. At an ℓ∞\ell_\infty budget of 0.03, it reports a 5.9-point macro-F1 gain over SDE-TGNN, with the same ordering across all three datasets and similar clean performance. The authors state that the certificate does not cover the evaluated encoder and eight-path predictor; no trained-model radii are reported.

Critical strengths
  1. The paper gives a useful certificate for the finite-step model. Theorem A and Corollary 1 turn model weight bounds and a mean score gap into a certified radius, accounting for sampling uncertainty. The argument allows low-rank noise. This is a concrete conditional result, even though it has not been applied to the trained detector.

  2. The reported PGD gain is substantial and consistent across the three datasets. At an infinity-norm budget of .03, macro-F1 is .931 versus .872 for SDE-TGNN, with similar clean performance. Five seeds and component ablations provide useful supporting evidence under the reported attack setup (Tables 2-4).

Critical weaknesses

The main robustness gain needs a more convincing attack check. The paper says PGD differentiates through the scored eight-path average, which is the appropriate starting point. It is unclear whether a meaningful advantage remains with more attack steps or restarts and appropriate handling of random paths on both models. Increasing the perturbation budget and running ablations with the same PGD do not answer this; Section 7 acknowledges that stronger adaptive results are unreported.

The provided code uses single-path logits in the shown PGD route (forward) and mean softmax probabilities for scoring (forward_mc). This differs from the eight-path mean-logit procedure described in Sections 4 and 6. Its relationship to the reported runs is unknown, so I would like the code version and settings used for Tables 2-4 clarified.

I find the finite-step certificate useful, but see the theoretical advance as modest on its own. The uncertainty in the main robustness result therefore keeps me on the weak-reject side. A clarified evaluation showing that a meaningful gain remains could change my recommendation.

Questions for authors
  1. Attack checks (evidence). At an ℓ∞\ell_\infty budget of 0.03, does the advantage over SDE-TGNN remain with more PGD steps or restarts on both models? Please explain whether paths are fixed or refreshed during attack and scoring, and how gradients are averaged where relevant. Existing convergence checks would help.

  2. Reported evaluation (clarification). Which code version and settings produced Tables 2-4? The provided code attacks single-path logits and scores averaged probabilities, while the paper describes an eight-path mean-logit predictor. Which procedure produced the reported results, and did any reported attack use BEL or ordinary backpropagation?

  3. Regularizer and path bound (clarification). Equation (4) uses top-2 gaps after averaging paths, whereas Proposition 2 uses signed single-path margins relative to a fixed class. What supports the claim that the regularizer reduces the first probability term in that bound? Is this a heuristic connection, or is there an argument or existing measurement linking them?

1 comment
AuthorsOpenReview ↗

Update after the five-seed re-run (see also the general update)

Thank you again for the detailed code check, which shaped this re-run.

Q1 (adaptive attacks). The full adaptive suite was run on SODE-Guard with EOT-matched adversarial training (module 2). Each attack is run once against fixed Brownian paths (seeds 0-7, a deterministic target) and once against fresh paths on every call:

  • PGD-40 and PGD-100 with restarts
  • APGD-CE and APGD-DLR with EOT
  • EOT-PGD (EOT 8)
  • a 64-path expected-margin attack
  • Square (black box)
  • transfer from two surrogates

Worst-case macro-F1 over all attacks (5 seeds):

  • fixed paths: 0.420 ± 0.010
  • fresh paths: 0.412 ± 0.012
  • without module 2: 0.245 (+0.174 with module 2, Holm-adjusted p = 0.016)

Individual attacks land between 0.42 and 0.45, so no single attack dominates.

The masking checks pass:

  • Varying eps: 0.462 at 0.01, 0.420 at 0.03, 0.290 at 0.1, 0.058 at 0.3, 0.004 at 1.0.
  • Varying steps: 0.425 at 10, 0.423 at 40, 0.427 at 100.
  • The black-box and transfer attacks are weaker than the white-box ones.

A full-strength re-check of seed 0 (PGD-200 with 5 restarts, EOT-32, a 256-path expected-margin attack, Square with 5,000 queries) is running now and will be reported in the revision.

Q2 (code version). There is now a single released pipeline covering data preparation, training, attacks, certification and aggregation: "https://anonymous.4open.science/r/SODE-ExtractGuard-Models/" though still under updates before deadline tomorrow.

  • What attacks differentiate through: the deployed predictor, i.e. the 8-path mean logits with the post-hoc logit adjustment. Every attack differentiates through exactly that predictor, by ordinary backpropagation through the Euler-Maruyama scheme.
  • Training randomness: Brownian seeds are fresh in every training step.
  • Data hygiene: duplicates and conflicting-label flows are removed before splitting.
  • Statistics: every number in the update is a mean over five seeds with a 95% CI. Comparisons use a paired bootstrap with Holm correction.

Q3 (regulariser and Proposition 2). As we reported, the regulariser had no measurable effect, and it is not part of the revised method. It is replaced by three modules, each with a significant measured effect:

  • imbalance-aware training: clean macro-F1 0.438 to 0.495;
  • EOT-matched adversarial training: worst case 0.245 to 0.420;
  • a smoothing certificate over the SDE's own noise: certified accuracy at l2 radius 0.273 from 0.096 (Theorem A) to 0.556.

Theorem A stays in the paper as a deterministic finite-step certificate. The headline certificate becomes the smoothing certificate, which samples input noise and Brownian paths jointly and is on par with a smoothed MLP (0.556 vs 0.555).

We hope this addresses your concerns, and would be grateful if you could take the update into account in your final assessment. Thanks.

Comments on the paper

3 comments
AuthorsOpenReview ↗

Update: five-seed re-run with three added modules (to all reviewers and the AC in particular)

Following the discussion, we added three modules to SODE-Guard and re-ran the full evaluation over five seeds with one released pipeline. The update addresses the three main concerns raised in the reviews: adaptive attacks, a certificate on a trained model, and code that matches the paper.

Protocol

  • Data: the container-flow part of ICS3D. Exact duplicates (325) and flows whose features appear with a different label (1,250) are removed before splitting, leaving 68,627 flows in 9 classes.
  • Splits: every seed draws a new stratified 60/20/20 split. Validation is halved into a selection half (early stopping, threshold tuning) and a held-out calibration half.
  • Attacks: l-inf eps = 0.03 (z-units) on 500 class-stratified test flows. They are PGD-40, PGD-100 with restarts, APGD-CE and APGD-DLR with EOT, EOT-PGD (EOT 8), a 64-path expected-margin attack, Square (black box) and transfer from two surrogates. We report the per-flow worst case over all attacks.
  • Statistics: mean ± 95% CI over 5 seeds. Every contrast is a paired bootstrap, Holm-corrected over a family declared before the runs.

The three modules (1) Imbalance-aware training: a logit-adjusted loss, a full training budget with early stopping on validation macro-F1, and a post-hoc logit adjustment tuned on validation data. (2) EOT-matched adversarial training: the training adversary differentiates through the deployed 8-path mean predictor with fresh Brownian paths. It uses N-FGSM, an eps ramp, and early stopping on robust validation macro-F1. (3) A randomised-smoothing certificate for the stochastic model: Gaussian input noise acts as a pure-diffusion prefix of the SDE. The certificate of Cohen et al. is applied over input noise and Brownian paths jointly; it is valid for [0,1]-valued base classifiers (Salman et al., 2019, Lemma 2).

Results (5 seeds)

  • Clean macro-F1: 0.438 ± 0.006 with the original recipe, 0.495 ± 0.009 with module 1 (+0.057, Holm-adjusted p = 0.016). The original recipe on the new split reproduces our rebuttal numbers (0.438 clean, 0.223 worst case), so the gains come from the modules.
  • Worst-case macro-F1: 0.245 before module 2, 0.420 ± 0.010 with it (+0.174, Holm-adjusted p = 0.016), at a clean cost of 0.027.
  • Certified accuracy at l2 radius 0.273 (this ball contains the l-inf 0.03 ball): 0.096 with Theorem A in our first reply, 0.556 ± 0.033 with module 3 on class-stratified flows, and 0.843 when reweighted to the test class mix.

Positioning

  • Robust neural baselines: trained with the same recipe, SODE-Guard is on par with the strongest of them. Worst case is 0.420 vs 0.424 for an adversarially trained MLP and 0.432 for an adversarially trained neural ODE; the differences (-0.001 and -0.005) are not significant. Certified accuracy is 0.556 vs 0.555 for a smoothed MLP.
  • Tree baselines: gradient-boosted trees and a random forest reach the highest clean macro-F1 (0.61-0.62) but fall to 0.02-0.04 under the black-box Square attack. Every adversarially trained neural model keeps 0.42-0.43.
  • This motivates robust neural detectors, and places SODE-Guard with these modules among the best of them.

Masking checks pass

  • F1 falls monotonically with eps, from 0.42 at eps = 0.03 to 0.004 at eps = 1.0.
  • More attack steps never help the model.
  • Square and transfer are weaker than the white-box attacks.
  • Fixed and fresh Brownian paths agree: 0.420 vs 0.412.

Already running, for the revision (i) A full-strength re-attack of seed 0 (PGD-200 with 5 restarts, EOT-32, a 256-path expected-margin attack, Square with 5,000 queries). (ii) Per-module ablations over five seeds. (iii) A per-class chronological split. Also planned: feature-constrained attacks and a second container dataset.

Code and configurations are in the updated anonymised repository. We thank the reviewers; this discussion has made the method markedly stronger. Thank you.

AuthorsOpenReview ↗

Revision summary: a three-module recipe makes SODE-Guard accurate on rare attacks, robust under adaptive attack and certified by its own randomness, validated over five seeds with a protocol built for security use.

We thank the reviewers and the Area Chair. Their comments shaped this revision, and the uploaded manuscript reflects all of the work carried out during the discussion.

What is new in the method (1) Imbalance-aware training: a logit-adjusted loss, a full training budget with early stopping on validation macro-F1, and a post-hoc logit adjustment tuned on a held-out selection split. (2) EOT-matched adversarial training: the training adversary attacks exactly the deployed eight-path predictor, with fresh Brownian paths at every step (N-FGSM, eps ramp, early stopping on robust validation macro-F1). (3) A smoothing certificate that uses the model's own noise. Proposition 1 proves that randomised smoothing remains sound when the base classifier is itself random. Gaussian input noise is a pure-diffusion prefix of the SDE, so SODE-Guard is certified by sampling input noise and Brownian paths jointly. Theorem A remains as a deterministic Lipschitz certificate for the discretised SDE.

How it is evaluated

  • Data: container flows of ICS3D, with duplicates and conflicting labels removed before splitting.
  • Seeds and splits: five seeds, each with a new 60/20/20 split.
  • Attacks: eight adaptive attacks (PGD with restarts, APGD-CE/DLR with EOT, EOT-PGD, an expected-margin attack, Square, transfer), each scored with fixed and with fresh Brownian paths.
  • Checks: gradient-masking checks, a full-strength re-attack, and a per-class chronological split.
  • Baselines: LightGBM, XGBoost and a random forest.
  • Statistics: paired-bootstrap contrasts with Holm correction, declared before the runs.

Main results (mean over five seeds)

  • Clean macro-F1: 0.438 -> 0.495 with module 1 (+0.057, Holm-adjusted p = 0.016).
  • Worst-case macro-F1 over all attacks: 0.245 -> 0.420 with module 2 (+0.174, Holm-adjusted p = 0.016). Fixed and fresh paths agree (0.420 vs 0.412).
  • Certified accuracy at l2 radius 0.273: 0.576 with module 3 (0.877 on the test class mix), and 0.491 at radius 0.5. The deterministic certificate gave 0.096 at radius 0.273.
  • SODE-Guard is on par with the strongest adversarially trained and smoothed baselines (MLP and neural ODE trained with the same recipe).
  • Gradient-boosted trees reach the best clean macro-F1 (0.62) but fall to 0.02-0.04 under a black-box attack, which shows where robust neural detectors are needed.
  • The full-strength re-attack and the chronological split confirm every conclusion.

How the revision answers each review

  • Reviewer vjJK:
    • Adaptive attacks are central to the evaluation, and the masking checks pass.
    • The released pipeline attacks and certifies exactly the predictor that is deployed.
    • The anti-concentration regulariser is replaced by three modules, each with a significant measured effect.
  • Reviewer FiAv:
    • The certificate is applied to trained detectors and is competitive with randomised smoothing of an MLP.
    • The stochastic model's distinctive strength is now explicit: its own paths are part of the smoothing distribution.
    • The protocol follows Arp et al. and TESSERACT: deduplication, multiple seeds, chronological evaluation, strong tree baselines.
    • The paper cites and discusses Liu et al. (CVPR 2020).
  • Reviewer kdck:
    • Every attack is scored with fixed and with fresh paths.
    • Recent robust and certified baselines are included, and BARS and MARS are discussed.
    • Figures 1-2 are redrawn around the method and the accuracy / robustness / certification trade-off.
    • The motivation is now grounded in the measured brittleness of the most accurate clean models.

Ablations Every module was ablated over five seeds:

  • the post-hoc adjustment and the training budget account for the clean gain;
  • single-step N-FGSM matches five-step PGD at lower cost;
  • the consistency weight trades certified accuracy at small radii against large radii.

Future work We will extend SODE-Guard along the lines the reviewers opened:

  • flow-sequence context for the rarest attack classes;
  • feature-constrained and packet-level attacks;
  • anisotropic noise in the certificate, in the spirit of BARS and MARS;
  • a second container corpus with cross-dataset transfer;
  • conformal alarm thresholds on the held-out calibration split.

The anonymised repository contains the code, the resumable multi-seed driver, a Colab notebook and every per-seed result file behind the tables and figures.

SODE-Guard and its open protocol give practitioners a ready path to intrusion detectors that are robust by measurement and certified by construction, and give researchers a reproducible benchmark for the next generation of robust detectors. Thank you

AuthorsOpenReview ↗

Update: under realisable attacks SODE-Guard keeps its robustness and meets a chosen false-alarm budget, while gradient-boosted trees remain brittle.

The updated PDF adds two analyses of the trained ICS3D models. Neither needed retraining.

  1. Feature-constrained attacks (new Table 6, Section 6.5). The attacker may change only the 44 client-side flow statistics. Server-side, port, protocol, flag and window features stay fixed, every value stays within its training range, and the 51 integer-valued features change by whole units. The suite is constrained PGD with restarts, EOT-PGD with fresh paths, a masked Square search and constrained transfer, over five seeds. Worst-case macro-F1:
  • SODE-Guard with adversarial training: 0.446, from 0.458 clean on the same flows.
  • MLP and neural ODE with adversarial training: 0.439 and 0.443.
  • Gradient-boosted trees and random forest: still 0.17-0.19, from about 0.60 clean. This answers the realism question raised by Reviewers FiAv and kdck: the robustness ranking holds when attacks must stay realisable.
  1. Conformal operating points (Section 6.5, Table 8). Alarm thresholds are fitted on the held-out calibration half.
  • The false-alarm budget holds on test flows for every model and seed: 0.007-0.011 at 1%, 0.047-0.050 at 5%.
  • At a 5% budget, SODE-Guard raises an alarm on 95.6% of attack flows, including 79-92% of the five rare attack classes.
  • Class-conditional conformal sets cover every class at least 80% of the time at the 90% level (SODE-Guard: 87%).
  1. Data provenance. The container flows are those of Amin et al. (AINA 2024), now cited. The attack classes are named in Table 9. The Node-RED remote-code-execution and container-escape stages share the network footprint of the reconnaissance stage, which explains why they are detected as attacks but rarely named.

The same protocol is running on two further flow corpora for the camera-ready version.

SODE-Guard now offers practitioners a detector that is robust under realistic attacks, certified by its own randomness and tunable to a false-alarm budget.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.